The tinyrobot community
Public experiments are open to everyone
Teach a little arm to reach. Turn noise into a path. See what a handful of pixels can tell a neural network. Explore local models and externally hosted experiences. Remix a reviewed walker experiment, run GPU inference, and share what you discover.
OPEN SOURCEtinyrobot / CartPole / browser experiment
A tiny brain. A whole lot of balance.
226 parameters learned to balance. Meet 128 tiny agents, switch between three stages of learning, and give them a gentle nudge.
Local inference on your device. Models load only after Start.
tinyrobot / MNIST / handwritten digits
Your handwriting. Ten possibilities.
Draw a digit and meet a tiny neural network trained on real MNIST handwriting. Compare its first guess with what it learned, entirely in your browser.
Local inference on your device. Models load only after Start.
tinyrobot / Image classification / Hugging Face Space
One image. One very specific question.
Explore image classification with a public Gradio demo by NimaBoscarino, featured in the Hugging Face embedding guide.
Externally hosted. Inputs are sent to the Space, not processed locally.
tinyrobot / Closed-loop simulation
Inspect an actual openpilot drive in MetaDrive.
Replay an operator-recorded, bounded openpilot simulation. Inspect actual camera observations, applied actions, lane position and inference diagnostics without launching a GPU.
Replay an operator-recorded GPU run. Public compute launch is not available.
tinyrobot / Control / Expert-policy distillation
A small motor. A learned sense of balance.
Watch a tiny neural torque controller swing a pendulum upright, then disturb its balance. Trained with synthetic expert-policy distillation in tinygrad.
Local inference on your device. Models load only after Start.
tinyrobot / Action chunking / synthetic robotics
Plan four moves. Take two. Think again.
Move a goal and watch a tiny learned policy push a cube across a tabletop. Peek inside its four-action plan and choose when it replans.
Local inference on your device. Models load only after Start.
tinyrobot / Robot reaching / behavior cloning
Two joints. One goal. A learned way there.
Move a target and watch a tiny imitation policy guide a two-link arm. Compare an untrained network with a policy learned from synthetic demonstrations.
Local inference on your device. Models load only after Start.
tinyrobot / Pretrained world model
Change the actions. Inspect the prediction.
A pretrained TD-MPC2 model imagines a walker’s future in latent space. Compare its reward forecasts with real MuJoCo simulation under exactly the same actions.
Replay an operator-recorded GPU run. Public compute launch is not available.
tinyrobot / Diffusion / trajectory sampling
A path takes shape, one denoising step at a time.
Choose two endpoints and a noise seed. Watch a tiny conditional diffusion network turn Gaussian noise into a smooth 2D path, entirely in your browser.
Local inference on your device. Models load only after Start.
tinyrobot / Visual servoing / browser experiment
Just pixels. No peeking.
A tiny learned visual encoder finds a red target in a synthetic camera. Move it, add distractions, and see exactly what the network sees.
Local inference on your device. Models load only after Start.
Small models, honest experiments
These are curated, source-available experiments, not a code playground. Public results need no account; reviewed GPU runs and private remixes require sign-in. Local models load only when you press Start experiment, then run on WebGPU or explicitly selected JavaScript CPU. The Hugging Face Space is hosted elsewhere and may process your inputs remotely.
Compare trained and untrained checkpoints, inspect the methods, and reproduce the results. Robotics scenes are synthetic teaching tasks, not validated robot controllers. Digit recognition uses MNIST; your handwriting may behave differently. No browser training, hardware control, or in-browser code editing.
Explore the project on GitHub